Beyond sentiment: knowledge graph-driven financial market forecasting with large language models-extracted enterprise relations and adaptive residual networks
Researchers have developed a new method for financial market forecasting by combining knowledge graphs, large language models, and adaptive deep learning. Their system, using T5-Relation-Aware Attention and BiLSTM with Attention and Layer-wise Residual Connections, improved sentiment classification F1-score by 16 percent and achieved an R² of 0.9695 and RMSE of 2.0861 in stock prediction. The lightweight knowledge graph schema enables efficient automation, but the model’s testing on only nine stocks limits generalizability.
What it examines
This paper introduces a new method for predicting financial markets by combining knowledge graphs, large language models, technical indicators, and deep learning. It aims to improve forecasting by capturing dynamic relationships between companies and integrating sentiment analysis from news, addressing limitations of traditional approaches.
What it concludes
The proposed model achieves higher accuracy in stock price prediction and sentiment classification, reducing errors and improving trend detection. Its practical applications include trading simulations and market analysis. The approach enables adaptive, low-cost knowledge graph construction and suggests future research in expanding generalization across markets and high-frequency trading scenarios.
Evidence objects
Researchers unveil a novel financial forecasting method, merging knowledge graphs, large language models, and adaptive deep learning, which dramatically boosts accuracy in sentiment analysis and stock prediction compared to traditional approaches.
key_findings bullet 1 · key_findings · validation V0
The systems lightweight knowledge graph schema, focusing on dynamic company relationships, enables efficient, automated construction and, with models like T5-RAA and BA-LRC, distinguishes positive and negative events based on enterprise ties.
key_findings bullet 2 · key_findings · validation V0
Results are striking: a 16% F1-score jump in sentiment classification and stock prediction with $R^2 = 0.9695$ and RMSE of 2.0861, though tested on only nine stocks, raising questions about generalizability.
key_findings bullet 3 · key_findings · validation V0
This paper presents a novel framework integrating LLM-extracted enterprise relations into a lightweight knowledge graph, enhancing sentiment analysis and time-series forecasting for financial markets. Its originality lies in automated KG construction via prompt engineering and the T5-Relation-Aware Attention model, yielding substantial predictive improvements. The integrations practical impact is compelling and significant.
key_findings bullet 4 · key_findings · validation V0
Raw abstract and provenance
- … The T5-RAA model proposed in this study, through … financial text sentiment analysis and relation extraction. … that our trading data is five-minute candlestick data and …
Source row: 308 · abstract type: snippet